Expose exact sample positions for raw model frames

[dissonance-git]

What changed

• Add model_output_frame_samples() to map unwrapped raw model frames back to their original-audio sample positions.
• Account for the mismatch between the input-window hop and the span represented by each retained output block.
• Validate overlap and hop arguments.
• Add deterministic tests covering the first window boundary and invalid alignment inputs.

Why

After overlapping output frames are removed, unwrap_output() concatenates the retained blocks. Treating the global frame index as index * FFT_HOP accumulates timing error because the default input hop is 36,164 samples while 142 retained frames span 36,352 samples.

For retained frame j in input window w, the exact original-audio position is:

w * hop_size + j * FFT_HOP

This does not alter inference arrays, note decoding, or MIDI output. It provides an opt-in alignment utility for consumers of raw note, onset, and contour arrays.

Prepared for upstream issue spotify#190.